Help center.
Find your way.

A simple way to turn a working set of documents into a useful canvas.

1. Choose a decision you need to make. Write it as a specific question: "Which customer problem should we investigate next?"
2. Add a small set of relevant sources rather than your entire archive. Give each source a recognizable title.
3. Connect the sources that belong together, then ask for a synthesis that separates evidence, assumptions and missing information.

The canvas prototype demonstrates this workflow. AI responses are currently illustrative, not live model inference.

Bring source material into view using the composer or canvas menu.

1. Use the plus button in the composer to choose files, or open the canvas menu and choose the file upload action.
2. Each file becomes a source card with its name, type and size. Move cards to group related material.
3. Small text, Markdown, CSV and JSON files can show a short preview. A card for a PDF or Office document does not yet mean its contents have been extracted.

Do not use the prototype as your only copy of a file. Document extraction, OCR and production storage are separate capabilities.

Pan, zoom and make room for your sources.

1. Drag a source card to reposition it. Use empty canvas space to navigate the working area.
2. Use the bottom-right zoom control to move in 5% steps between 5% and 200%. The canvas starts at 90%.
3. Press Ctrl+U on Windows or Cmd+U on macOS to hide or restore the composer when you want an unobstructed canvas.

Keyboard shortcuts apply to the canvas app, not this website.

An arrow should explain why two pieces of context belong together.

1. Connect supporting evidence to a question or synthesis step rather than drawing every possible relationship.
2. Keep conflicting sources visible. A connection shows relevance, not proof that a claim is true.
3. Review the source behind an important conclusion before carrying it into a brief or decision.

A visual graph helps you organize reasoning; it does not replace checking the original material.

Low, Medium and High express how much reasoning a task deserves.

1. Low suits a narrow question or a first pass over a small working set.
2. Medium is the default for everyday synthesis and comparing a few sources.
3. High is intended for more involved comparisons and questions with competing evidence. Make the question specific before increasing effort.

The selector exists in the prototype. Live inference, measured usage and production quotas are not connected yet.

Supervised, Auto-accept edits, Auto and Full access describe approval preferences.

1. Supervised asks before commands and changes. Auto-accept edits approves edits while asking before other actions.
2. Auto is intended to approve routine actions where the provider supports that behavior. Full access expresses a preference for actions without approval prompts.
3. Start with Supervised when evaluating a workflow, and review any action that could affect another system.

These are prototype UI settings, not a deployed tool-permission system. They do not currently authorize connected-app actions.

Keep the original material and any important conclusions outside the prototype.

1. Do not assume that a canvas is backed up or synchronized to an account. Durable cloud workspace storage is not implemented in the current prototype.
2. Keep your original files and record important notes in a reliable location before refreshing.
3. If a view becomes confusing, check the zoom setting and reopen the composer before starting over.

Production backup and recovery are required before accepting customer data; this is not a promise that prototype sessions can be recovered.

A good report explains the steps, expected result and what actually happened.

1. Include your browser or operating system and the sequence of actions that triggered the problem.
2. Use a small, non-confidential sample if the issue involves a file. Remove customer names, tokens and personal information from screenshots.
3. Send the report by email. For a potential vulnerability, follow the security reporting guidance instead of posting details publicly.

We do not promise a response time or recovery of unsaved prototype work.